Message ID | 20190128212437.11597-1-dennis@kernel.org (mailing list archive) |
---|---|
Headers | show |
Series | btrfs: add zstd compression level support | expand |
On Mon, Jan 28, 2019 at 04:24:26PM -0500, Dennis Zhou wrote: > As mentioned above, a requirement that differs zstd from zlib is that > higher levels of compression require more memory. To manage this, each > compression level has its own queue of workspaces. A global LRU is used > to help with reclaim. To guarantee forward progress, a max level > workspace is preallocated and hidden from the LRU. Here I'd like to bring up what was mentioned in previous iteration, the workspace sizes. Level Compression Memory 1 0.8 MB 2 1.0 MB 3 1.3 MB 4 0.9 MB 5 1.4 MB 6 1.5 MB 7 1.4 MB 8 1.8 MB 9 1.8 MB 10 1.8 MB 11 1.8 MB 12 1.8 MB 13 2.3 MB 14 2.6 MB 15 2.6 MB and decompression needs memory of level 1. The sizes can be grouped together to say 3 sizes, I'm not sure that we'd really need 15 distinct workspaces. The reclaim mechanism helps, but I'd rather keep a smaller number of workspaces that covers average use. Default level is 3, that's 1.3 MiB, that also covers level 1, 2 and 4. For 5 to 12 it's 1.8 and the rest is 2.6 MiB. > btrfs filesystem 10 times and then read back after dropping the caches. > The btrfs filesystem was on an SSD. > > Level Ratio Compression (MB/s) Decompression (MB/s) > 1 2.658 438.47 910.51 > 2 2.744 364.86 886.55 > 3 2.801 336.33 828.41 > 4 2.858 286.71 886.55 > 5 2.916 212.77 556.84 > 6 2.363 119.82 990.85 > 7 3.000 154.06 849.30 > 8 3.011 159.54 875.03 > 9 3.025 100.51 940.15 > 10 3.033 118.97 616.26 > 11 3.036 94.19 802.11 > 12 3.037 73.45 931.49 > 13 3.041 55.17 835.26 > 14 3.087 44.70 716.78 > 15 3.126 37.30 878.84 From my casual user's perspective, I'd use the level 1 for speed, 7 for better ratio and 15 for the best compression. Anything else does not look good, though the results would vary based on the data set. I assume that the silesia corpus serves as a good approximation of the worst case average. The levels 7-14 strike particularly obvious pattern: same ratio but the speed gets worse with each level. Taking the default level into account, (my) recommended levels would be 1, 3, 7, 15. I went through the patches, looks mostly ok, I don't like the indirections but at the moment it's an implementation detail as I'd like to agree on the overall approach first. We might need a few revisions or cleanup rounds to converge to an efficient solution, the advantage here is that it's all in-memory and without compatibility concerns once the level support for zstd is in and works. For that reason, I'm not opposed to the current version of the patchset. Given the time in development schedule, it's really close to code freeze, but the functionality has a narrow scope so I'm tentatively counting with it for 5.1.
> On Jan 29, 2019, at 9:18 AM, David Sterba <dsterba@suse.cz> wrote: > > On Mon, Jan 28, 2019 at 04:24:26PM -0500, Dennis Zhou wrote: >> As mentioned above, a requirement that differs zstd from zlib is that >> higher levels of compression require more memory. To manage this, each >> compression level has its own queue of workspaces. A global LRU is used >> to help with reclaim. To guarantee forward progress, a max level >> workspace is preallocated and hidden from the LRU. > > Here I'd like to bring up what was mentioned in previous iteration, the > workspace sizes. > > Level Compression Memory > 1 0.8 MB > 2 1.0 MB > 3 1.3 MB > 4 0.9 MB > 5 1.4 MB > 6 1.5 MB > 7 1.4 MB > 8 1.8 MB > 9 1.8 MB > 10 1.8 MB > 11 1.8 MB > 12 1.8 MB > 13 2.3 MB > 14 2.6 MB > 15 2.6 MB > > and decompression needs memory of level 1. The sizes can be grouped > together to say 3 sizes, I'm not sure that we'd really need 15 distinct > workspaces. The reclaim mechanism helps, but I'd rather keep a smaller > number of workspaces that covers average use. > > Default level is 3, that's 1.3 MiB, that also covers level 1, 2 and 4. > For 5 to 12 it's 1.8 and the rest is 2.6 MiB. > >> btrfs filesystem 10 times and then read back after dropping the caches. >> The btrfs filesystem was on an SSD. >> >> Level Ratio Compression (MB/s) Decompression (MB/s) >> 1 2.658 438.47 910.51 >> 2 2.744 364.86 886.55 >> 3 2.801 336.33 828.41 >> 4 2.858 286.71 886.55 >> 5 2.916 212.77 556.84 >> 6 2.363 119.82 990.85 >> 7 3.000 154.06 849.30 >> 8 3.011 159.54 875.03 >> 9 3.025 100.51 940.15 >> 10 3.033 118.97 616.26 >> 11 3.036 94.19 802.11 >> 12 3.037 73.45 931.49 >> 13 3.041 55.17 835.26 >> 14 3.087 44.70 716.78 >> 15 3.126 37.30 878.84 > > From my casual user's perspective, I'd use the level 1 for speed, 7 for > better ratio and 15 for the best compression. Anything else does not > look good, though the results would vary based on the data set. I > assume that the silesia corpus serves as a good approximation of the > worst case average. > > The levels 7-14 strike particularly obvious pattern: same ratio but the > speed gets worse with each level. Taking the default level into account, > (my) recommended levels would be 1, 3, 7, 15. Silesia is used because it is a standard corpus, and I'd call it about average, but there is a lot of variance and extreme edge case data. The intermediate strategies will change in effectiveness on different types of data. For example, the lower levels are generally more effective on text, and you want slightly higher levels for non-text data, because they can find shorter matches. Upstream zstd also shifts around its levels, and the memory usage of each level from time-to-time, and I am going to update zstd in the kernel in this next year, since we are slowing down development. The shifts will be small though. It could make sense to map the levels into size classes, since that could reduce memory spikes, at the cost of higher stead-state memory usage. I'm not familiar with the machinery used in these patches, so I can't actually say much. I would probably use levels 1, 3, 7 (after it is made monotonic), 12, and 15. You might skip 7, but leave 12. > I went through the patches, looks mostly ok, I don't like the > indirections but at the moment it's an implementation detail as I'd like > to agree on the overall approach first. > > We might need a few revisions or cleanup rounds to converge to an > efficient solution, the advantage here is that it's all in-memory and > without compatibility concerns once the level support for zstd is in and > works. > > For that reason, I'm not opposed to the current version of the patchset. > Given the time in development schedule, it's really close to code > freeze, but the functionality has a narrow scope so I'm tentatively > counting with it for 5.1.
Hi David, On Tue, Jan 29, 2019 at 06:18:30PM +0100, David Sterba wrote: > On Mon, Jan 28, 2019 at 04:24:26PM -0500, Dennis Zhou wrote: > > As mentioned above, a requirement that differs zstd from zlib is that > > higher levels of compression require more memory. To manage this, each > > compression level has its own queue of workspaces. A global LRU is used > > to help with reclaim. To guarantee forward progress, a max level > > workspace is preallocated and hidden from the LRU. > > Here I'd like to bring up what was mentioned in previous iteration, the > workspace sizes. > > Level Compression Memory > 1 0.8 MB > 2 1.0 MB > 3 1.3 MB > 4 0.9 MB > 5 1.4 MB > 6 1.5 MB > 7 1.4 MB > 8 1.8 MB > 9 1.8 MB > 10 1.8 MB > 11 1.8 MB > 12 1.8 MB > 13 2.3 MB > 14 2.6 MB > 15 2.6 MB > > and decompression needs memory of level 1. The sizes can be grouped > together to say 3 sizes, I'm not sure that we'd really need 15 distinct > workspaces. The reclaim mechanism helps, but I'd rather keep a smaller > number of workspaces that covers average use. > > Default level is 3, that's 1.3 MiB, that also covers level 1, 2 and 4. > For 5 to 12 it's 1.8 and the rest is 2.6 MiB. > I realize the current implementation doesn't have a monotonic memory requirement guarantee. I've added that, and below is updated memory requirements per level. I've updated the commit to include this too. Level Memory (KB) 1 780 2 1004 3 1260 4 1260 5 1388 6 1516 7 1516 8 1772 9 1772 10 1772 11 1772 12 1772 13 2284 14 2547 15 2547 > > btrfs filesystem 10 times and then read back after dropping the caches. > > The btrfs filesystem was on an SSD. > > > > Level Ratio Compression (MB/s) Decompression (MB/s) > > 1 2.658 438.47 910.51 > > 2 2.744 364.86 886.55 > > 3 2.801 336.33 828.41 > > 4 2.858 286.71 886.55 > > 5 2.916 212.77 556.84 > > 6 2.363 119.82 990.85 > > 7 3.000 154.06 849.30 > > 8 3.011 159.54 875.03 > > 9 3.025 100.51 940.15 > > 10 3.033 118.97 616.26 > > 11 3.036 94.19 802.11 > > 12 3.037 73.45 931.49 > > 13 3.041 55.17 835.26 > > 14 3.087 44.70 716.78 > > 15 3.126 37.30 878.84 > > From my casual user's perspective, I'd use the level 1 for speed, 7 for > better ratio and 15 for the best compression. Anything else does not > look good, though the results would vary based on the data set. I > assume that the silesia corpus serves as a good approximation of the > worst case average. > > The levels 7-14 strike particularly obvious pattern: same ratio but the > speed gets worse with each level. Taking the default level into account, > (my) recommended levels would be 1, 3, 7, 15. > I do see why we want to limit the number of levels as the memory requirements do kind of bucket themselves. But, this means when zstd gets updated, we'd have to reevaluate the compression levels btrfs supports. I'm not sure it's a great idea to have that dependency. I imagine we could offer some level of guidance, but it really would be up to the user to figure out what works best for them. The reclaim mechanism only keeps workspaces around if they are being used by the appropriate level. So, the memory overhead is actively used memory and if not, it is reclaimed after at most ~2 minutes later. I also scan up before allocating a workspace, so that should help limit the number of workspaces in circulation. > I went through the patches, looks mostly ok, I don't like the > indirections but at the moment it's an implementation detail as I'd like > to agree on the overall approach first. > > We might need a few revisions or cleanup rounds to converge to an > efficient solution, the advantage here is that it's all in-memory and > without compatibility concerns once the level support for zstd is in and > works. > > For that reason, I'm not opposed to the current version of the patchset. > Given the time in development schedule, it's really close to code > freeze, but the functionality has a narrow scope so I'm tentatively > counting with it for 5.1. Thanks, Dennis
On Wed, Jan 30, 2019 at 12:40:59PM -0500, Dennis Zhou wrote: > Hi David, > > On Tue, Jan 29, 2019 at 06:18:30PM +0100, David Sterba wrote: > > On Mon, Jan 28, 2019 at 04:24:26PM -0500, Dennis Zhou wrote: > > > As mentioned above, a requirement that differs zstd from zlib is that > > > higher levels of compression require more memory. To manage this, each > > > compression level has its own queue of workspaces. A global LRU is used > > > to help with reclaim. To guarantee forward progress, a max level > > > workspace is preallocated and hidden from the LRU. > > > > Here I'd like to bring up what was mentioned in previous iteration, the > > workspace sizes. > > > > Level Compression Memory > > 1 0.8 MB > > 2 1.0 MB > > 3 1.3 MB > > 4 0.9 MB > > 5 1.4 MB > > 6 1.5 MB > > 7 1.4 MB > > 8 1.8 MB > > 9 1.8 MB > > 10 1.8 MB > > 11 1.8 MB > > 12 1.8 MB > > 13 2.3 MB > > 14 2.6 MB > > 15 2.6 MB > > > > and decompression needs memory of level 1. The sizes can be grouped > > together to say 3 sizes, I'm not sure that we'd really need 15 distinct > > workspaces. The reclaim mechanism helps, but I'd rather keep a smaller > > number of workspaces that covers average use. > > > > Default level is 3, that's 1.3 MiB, that also covers level 1, 2 and 4. > > For 5 to 12 it's 1.8 and the rest is 2.6 MiB. > > > > I realize the current implementation doesn't have a monotonic memory > requirement guarantee. I've added that, and below is updated memory > requirements per level. I've updated the commit to include this too. > > Level Memory (KB) > 1 780 > 2 1004 > 3 1260 > 4 1260 > 5 1388 > 6 1516 > 7 1516 > 8 1772 > 9 1772 > 10 1772 > 11 1772 > 12 1772 > 13 2284 > 14 2547 > 15 2547 > > > > btrfs filesystem 10 times and then read back after dropping the caches. > > > The btrfs filesystem was on an SSD. > > > > > > Level Ratio Compression (MB/s) Decompression (MB/s) > > > 1 2.658 438.47 910.51 > > > 2 2.744 364.86 886.55 > > > 3 2.801 336.33 828.41 > > > 4 2.858 286.71 886.55 > > > 5 2.916 212.77 556.84 > > > 6 2.363 119.82 990.85 > > > 7 3.000 154.06 849.30 > > > 8 3.011 159.54 875.03 > > > 9 3.025 100.51 940.15 > > > 10 3.033 118.97 616.26 > > > 11 3.036 94.19 802.11 > > > 12 3.037 73.45 931.49 > > > 13 3.041 55.17 835.26 > > > 14 3.087 44.70 716.78 > > > 15 3.126 37.30 878.84 > > > > From my casual user's perspective, I'd use the level 1 for speed, 7 for > > better ratio and 15 for the best compression. Anything else does not > > look good, though the results would vary based on the data set. I > > assume that the silesia corpus serves as a good approximation of the > > worst case average. > > > > The levels 7-14 strike particularly obvious pattern: same ratio but the > > speed gets worse with each level. Taking the default level into account, > > (my) recommended levels would be 1, 3, 7, 15. > > > > I do see why we want to limit the number of levels as the memory > requirements do kind of bucket themselves. But, this means when zstd > gets updated, we'd have to reevaluate the compression levels btrfs > supports. I'm not sure it's a great idea to have that dependency. > I imagine we could offer some level of guidance, but it really would be > up to the user to figure out what works best for them. If it was not clear, I did not mean to have only 4 levels, keep all 15 same as there are 9 for zlib. The guildelines would be desirable and I don't want to make decision for the user which level to pick. So we don't disagree. > The reclaim mechanism only keeps workspaces around if they are being > used by the appropriate level. So, the memory overhead is actively used > memory and if not, it is reclaimed after at most ~2 minutes later. I > also scan up before allocating a workspace, so that should help limit > the number of workspaces in circulation. We'd need to observe that in practice before doing refinements, simpler logic is better for the start. There's some penalty caused by the allocation if there are no workspaces at all, as the amount of memory is quite large for kernel. This could stress the memory subsystem also because the memory has to be either contiguous or vmalloced. As the memory is released soon, all the work might need to be done again and again. So, more than one preallocated workspace could be good but the number of levels does not make it easy to choose which one.
On Thu, Jan 31, 2019 at 03:04:36PM +0100, David Sterba wrote: > On Wed, Jan 30, 2019 at 12:40:59PM -0500, Dennis Zhou wrote: > > Hi David, > > > > On Tue, Jan 29, 2019 at 06:18:30PM +0100, David Sterba wrote: > > > On Mon, Jan 28, 2019 at 04:24:26PM -0500, Dennis Zhou wrote: > > > > As mentioned above, a requirement that differs zstd from zlib is that > > > > higher levels of compression require more memory. To manage this, each > > > > compression level has its own queue of workspaces. A global LRU is used > > > > to help with reclaim. To guarantee forward progress, a max level > > > > workspace is preallocated and hidden from the LRU. > > > > > > Here I'd like to bring up what was mentioned in previous iteration, the > > > workspace sizes. > > > > > > Level Compression Memory > > > 1 0.8 MB > > > 2 1.0 MB > > > 3 1.3 MB > > > 4 0.9 MB > > > 5 1.4 MB > > > 6 1.5 MB > > > 7 1.4 MB > > > 8 1.8 MB > > > 9 1.8 MB > > > 10 1.8 MB > > > 11 1.8 MB > > > 12 1.8 MB > > > 13 2.3 MB > > > 14 2.6 MB > > > 15 2.6 MB > > > > > > and decompression needs memory of level 1. The sizes can be grouped > > > together to say 3 sizes, I'm not sure that we'd really need 15 distinct > > > workspaces. The reclaim mechanism helps, but I'd rather keep a smaller > > > number of workspaces that covers average use. > > > > > > Default level is 3, that's 1.3 MiB, that also covers level 1, 2 and 4. > > > For 5 to 12 it's 1.8 and the rest is 2.6 MiB. > > > > > > > I realize the current implementation doesn't have a monotonic memory > > requirement guarantee. I've added that, and below is updated memory > > requirements per level. I've updated the commit to include this too. > > > > Level Memory (KB) > > 1 780 > > 2 1004 > > 3 1260 > > 4 1260 > > 5 1388 > > 6 1516 > > 7 1516 > > 8 1772 > > 9 1772 > > 10 1772 > > 11 1772 > > 12 1772 > > 13 2284 > > 14 2547 > > 15 2547 > > > > > > btrfs filesystem 10 times and then read back after dropping the caches. > > > > The btrfs filesystem was on an SSD. > > > > > > > > Level Ratio Compression (MB/s) Decompression (MB/s) > > > > 1 2.658 438.47 910.51 > > > > 2 2.744 364.86 886.55 > > > > 3 2.801 336.33 828.41 > > > > 4 2.858 286.71 886.55 > > > > 5 2.916 212.77 556.84 > > > > 6 2.363 119.82 990.85 > > > > 7 3.000 154.06 849.30 > > > > 8 3.011 159.54 875.03 > > > > 9 3.025 100.51 940.15 > > > > 10 3.033 118.97 616.26 > > > > 11 3.036 94.19 802.11 > > > > 12 3.037 73.45 931.49 > > > > 13 3.041 55.17 835.26 > > > > 14 3.087 44.70 716.78 > > > > 15 3.126 37.30 878.84 > > > > > > From my casual user's perspective, I'd use the level 1 for speed, 7 for > > > better ratio and 15 for the best compression. Anything else does not > > > look good, though the results would vary based on the data set. I > > > assume that the silesia corpus serves as a good approximation of the > > > worst case average. > > > > > > The levels 7-14 strike particularly obvious pattern: same ratio but the > > > speed gets worse with each level. Taking the default level into account, > > > (my) recommended levels would be 1, 3, 7, 15. > > > > > > > I do see why we want to limit the number of levels as the memory > > requirements do kind of bucket themselves. But, this means when zstd > > gets updated, we'd have to reevaluate the compression levels btrfs > > supports. I'm not sure it's a great idea to have that dependency. > > I imagine we could offer some level of guidance, but it really would be > > up to the user to figure out what works best for them. > > If it was not clear, I did not mean to have only 4 levels, keep all 15 > same as there are 9 for zlib. The guildelines would be desirable and I > don't want to make decision for the user which level to pick. So we > don't disagree. > I see, that was my misunderstanding. > > The reclaim mechanism only keeps workspaces around if they are being > > used by the appropriate level. So, the memory overhead is actively used > > memory and if not, it is reclaimed after at most ~2 minutes later. I > > also scan up before allocating a workspace, so that should help limit > > the number of workspaces in circulation. > > We'd need to observe that in practice before doing refinements, simpler > logic is better for the start. > > There's some penalty caused by the allocation if there are no workspaces > at all, as the amount of memory is quite large for kernel. > This could stress the memory subsystem also because the memory has to be > either contiguous or vmalloced. As the memory is released soon, all the > work might need to be done again and again. So, more than one > preallocated workspace could be good but the number of levels does not > make it easy to choose which one. That makes sense. I don't have an answer for how to balance the number of workspaces, but am happy to iterate on this as we get more data. If no one has any other comments on the series after another day or so I can send v2 with the handful of things people have mentioned and the monotonic memory requirement patch. Thanks, Dennis